1 /* 2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the BSD-type 8 * license below: 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 17 * Redistributions in binary form must reproduce the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer in the documentation and/or other materials provided 20 * with the distribution. 21 * 22 * Neither the name of the Network Appliance, Inc. nor the names of 23 * its contributors may be used to endorse or promote products 24 * derived from this software without specific prior written 25 * permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 38 */ 39 40 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H 41 #define _LINUX_SUNRPC_XPRT_RDMA_H 42 43 #include <linux/wait.h> /* wait_queue_head_t, etc */ 44 #include <linux/spinlock.h> /* spinlock_t, etc */ 45 #include <linux/atomic.h> /* atomic_t, etc */ 46 #include <linux/workqueue.h> /* struct work_struct */ 47 48 #include <rdma/rdma_cm.h> /* RDMA connection api */ 49 #include <rdma/ib_verbs.h> /* RDMA verbs api */ 50 51 #include <linux/sunrpc/clnt.h> /* rpc_xprt */ 52 #include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */ 53 #include <linux/sunrpc/xprtrdma.h> /* xprt parameters */ 54 55 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */ 56 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */ 57 58 #define RPCRDMA_BIND_TO (60U * HZ) 59 #define RPCRDMA_INIT_REEST_TO (5U * HZ) 60 #define RPCRDMA_MAX_REEST_TO (30U * HZ) 61 #define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ) 62 63 /* 64 * Interface Adapter -- one per transport instance 65 */ 66 struct rpcrdma_ia { 67 const struct rpcrdma_memreg_ops *ri_ops; 68 rwlock_t ri_qplock; 69 struct ib_device *ri_device; 70 struct rdma_cm_id *ri_id; 71 struct ib_pd *ri_pd; 72 struct ib_mr *ri_dma_mr; 73 struct completion ri_done; 74 int ri_async_rc; 75 unsigned int ri_max_frmr_depth; 76 struct ib_qp_attr ri_qp_attr; 77 struct ib_qp_init_attr ri_qp_init_attr; 78 }; 79 80 /* 81 * RDMA Endpoint -- one per transport instance 82 */ 83 84 struct rpcrdma_ep { 85 atomic_t rep_cqcount; 86 int rep_cqinit; 87 int rep_connected; 88 struct ib_qp_init_attr rep_attr; 89 wait_queue_head_t rep_connect_wait; 90 struct rdma_conn_param rep_remote_cma; 91 struct sockaddr_storage rep_remote_addr; 92 struct delayed_work rep_connect_worker; 93 }; 94 95 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit) 96 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount) 97 98 /* Force completion handler to ignore the signal 99 */ 100 #define RPCRDMA_IGNORE_COMPLETION (0ULL) 101 102 /* Pre-allocate extra Work Requests for handling backward receives 103 * and sends. This is a fixed value because the Work Queues are 104 * allocated when the forward channel is set up. 105 */ 106 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 107 #define RPCRDMA_BACKWARD_WRS (8) 108 #else 109 #define RPCRDMA_BACKWARD_WRS (0) 110 #endif 111 112 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV 113 * 114 * The below structure appears at the front of a large region of kmalloc'd 115 * memory, which always starts on a good alignment boundary. 116 */ 117 118 struct rpcrdma_regbuf { 119 size_t rg_size; 120 struct rpcrdma_req *rg_owner; 121 struct ib_sge rg_iov; 122 __be32 rg_base[0] __attribute__ ((aligned(256))); 123 }; 124 125 static inline u64 126 rdmab_addr(struct rpcrdma_regbuf *rb) 127 { 128 return rb->rg_iov.addr; 129 } 130 131 static inline u32 132 rdmab_length(struct rpcrdma_regbuf *rb) 133 { 134 return rb->rg_iov.length; 135 } 136 137 static inline u32 138 rdmab_lkey(struct rpcrdma_regbuf *rb) 139 { 140 return rb->rg_iov.lkey; 141 } 142 143 static inline struct rpcrdma_msg * 144 rdmab_to_msg(struct rpcrdma_regbuf *rb) 145 { 146 return (struct rpcrdma_msg *)rb->rg_base; 147 } 148 149 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN) 150 151 /* 152 * struct rpcrdma_rep -- this structure encapsulates state required to recv 153 * and complete a reply, asychronously. It needs several pieces of 154 * state: 155 * o recv buffer (posted to provider) 156 * o ib_sge (also donated to provider) 157 * o status of reply (length, success or not) 158 * o bookkeeping state to get run by tasklet (list, etc) 159 * 160 * These are allocated during initialization, per-transport instance; 161 * however, the tasklet execution list itself is global, as it should 162 * always be pretty short. 163 * 164 * N of these are associated with a transport instance, and stored in 165 * struct rpcrdma_buffer. N is the max number of outstanding requests. 166 */ 167 168 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE) 169 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */ 170 171 struct rpcrdma_buffer; 172 173 struct rpcrdma_rep { 174 unsigned int rr_len; 175 struct ib_device *rr_device; 176 struct rpcrdma_xprt *rr_rxprt; 177 struct work_struct rr_work; 178 struct list_head rr_list; 179 struct rpcrdma_regbuf *rr_rdmabuf; 180 }; 181 182 #define RPCRDMA_BAD_LEN (~0U) 183 184 /* 185 * struct rpcrdma_mw - external memory region metadata 186 * 187 * An external memory region is any buffer or page that is registered 188 * on the fly (ie, not pre-registered). 189 * 190 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During 191 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in 192 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep 193 * track of registration metadata while each RPC is pending. 194 * rpcrdma_deregister_external() uses this metadata to unmap and 195 * release these resources when an RPC is complete. 196 */ 197 enum rpcrdma_frmr_state { 198 FRMR_IS_INVALID, /* ready to be used */ 199 FRMR_IS_VALID, /* in use */ 200 FRMR_IS_STALE, /* failed completion */ 201 }; 202 203 struct rpcrdma_frmr { 204 struct scatterlist *sg; 205 int sg_nents; 206 struct ib_mr *fr_mr; 207 enum rpcrdma_frmr_state fr_state; 208 struct work_struct fr_work; 209 struct rpcrdma_xprt *fr_xprt; 210 bool fr_waiter; 211 struct completion fr_linv_done;; 212 union { 213 struct ib_reg_wr fr_regwr; 214 struct ib_send_wr fr_invwr; 215 }; 216 }; 217 218 struct rpcrdma_fmr { 219 struct ib_fmr *fmr; 220 u64 *physaddrs; 221 }; 222 223 struct rpcrdma_mw { 224 union { 225 struct rpcrdma_fmr fmr; 226 struct rpcrdma_frmr frmr; 227 } r; 228 void (*mw_sendcompletion)(struct ib_wc *); 229 struct list_head mw_list; 230 struct list_head mw_all; 231 }; 232 233 /* 234 * struct rpcrdma_req -- structure central to the request/reply sequence. 235 * 236 * N of these are associated with a transport instance, and stored in 237 * struct rpcrdma_buffer. N is the max number of outstanding requests. 238 * 239 * It includes pre-registered buffer memory for send AND recv. 240 * The recv buffer, however, is not owned by this structure, and 241 * is "donated" to the hardware when a recv is posted. When a 242 * reply is handled, the recv buffer used is given back to the 243 * struct rpcrdma_req associated with the request. 244 * 245 * In addition to the basic memory, this structure includes an array 246 * of iovs for send operations. The reason is that the iovs passed to 247 * ib_post_{send,recv} must not be modified until the work request 248 * completes. 249 * 250 * NOTES: 251 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we 252 * marshal. The number needed varies depending on the iov lists that 253 * are passed to us, the memory registration mode we are in, and if 254 * physical addressing is used, the layout. 255 */ 256 257 struct rpcrdma_mr_seg { /* chunk descriptors */ 258 struct rpcrdma_mw *rl_mw; /* registered MR */ 259 u64 mr_base; /* registration result */ 260 u32 mr_rkey; /* registration result */ 261 u32 mr_len; /* length of chunk or segment */ 262 int mr_nsegs; /* number of segments in chunk or 0 */ 263 enum dma_data_direction mr_dir; /* segment mapping direction */ 264 dma_addr_t mr_dma; /* segment mapping address */ 265 size_t mr_dmalen; /* segment mapping length */ 266 struct page *mr_page; /* owning page, if any */ 267 char *mr_offset; /* kva if no page, else offset */ 268 }; 269 270 #define RPCRDMA_MAX_IOVS (2) 271 272 struct rpcrdma_req { 273 struct list_head rl_free; 274 unsigned int rl_niovs; 275 unsigned int rl_nchunks; 276 unsigned int rl_connect_cookie; 277 struct rpcrdma_buffer *rl_buffer; 278 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */ 279 struct ib_sge rl_send_iov[RPCRDMA_MAX_IOVS]; 280 struct rpcrdma_regbuf *rl_rdmabuf; 281 struct rpcrdma_regbuf *rl_sendbuf; 282 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS]; 283 284 struct list_head rl_all; 285 bool rl_backchannel; 286 }; 287 288 static inline struct rpcrdma_req * 289 rpcr_to_rdmar(struct rpc_rqst *rqst) 290 { 291 void *buffer = rqst->rq_buffer; 292 struct rpcrdma_regbuf *rb; 293 294 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base); 295 return rb->rg_owner; 296 } 297 298 /* 299 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for 300 * inline requests/replies, and client/server credits. 301 * 302 * One of these is associated with a transport instance 303 */ 304 struct rpcrdma_buffer { 305 spinlock_t rb_mwlock; /* protect rb_mws list */ 306 struct list_head rb_mws; 307 struct list_head rb_all; 308 char *rb_pool; 309 310 spinlock_t rb_lock; /* protect buf lists */ 311 struct list_head rb_send_bufs; 312 struct list_head rb_recv_bufs; 313 u32 rb_max_requests; 314 315 u32 rb_bc_srv_max_requests; 316 spinlock_t rb_reqslock; /* protect rb_allreqs */ 317 struct list_head rb_allreqs; 318 319 u32 rb_bc_max_requests; 320 }; 321 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia) 322 323 /* 324 * Internal structure for transport instance creation. This 325 * exists primarily for modularity. 326 * 327 * This data should be set with mount options 328 */ 329 struct rpcrdma_create_data_internal { 330 struct sockaddr_storage addr; /* RDMA server address */ 331 unsigned int max_requests; /* max requests (slots) in flight */ 332 unsigned int rsize; /* mount rsize - max read hdr+data */ 333 unsigned int wsize; /* mount wsize - max write hdr+data */ 334 unsigned int inline_rsize; /* max non-rdma read data payload */ 335 unsigned int inline_wsize; /* max non-rdma write data payload */ 336 unsigned int padding; /* non-rdma write header padding */ 337 }; 338 339 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \ 340 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize) 341 342 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\ 343 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize) 344 345 #define RPCRDMA_INLINE_PAD_VALUE(rq)\ 346 rpcx_to_rdmad(rq->rq_xprt).padding 347 348 /* 349 * Statistics for RPCRDMA 350 */ 351 struct rpcrdma_stats { 352 unsigned long read_chunk_count; 353 unsigned long write_chunk_count; 354 unsigned long reply_chunk_count; 355 356 unsigned long long total_rdma_request; 357 unsigned long long total_rdma_reply; 358 359 unsigned long long pullup_copy_count; 360 unsigned long long fixup_copy_count; 361 unsigned long hardway_register_count; 362 unsigned long failed_marshal_count; 363 unsigned long bad_reply_count; 364 unsigned long nomsg_call_count; 365 unsigned long bcall_count; 366 }; 367 368 /* 369 * Per-registration mode operations 370 */ 371 struct rpcrdma_xprt; 372 struct rpcrdma_memreg_ops { 373 int (*ro_map)(struct rpcrdma_xprt *, 374 struct rpcrdma_mr_seg *, int, bool); 375 void (*ro_unmap_sync)(struct rpcrdma_xprt *, 376 struct rpcrdma_req *); 377 int (*ro_unmap)(struct rpcrdma_xprt *, 378 struct rpcrdma_mr_seg *); 379 int (*ro_open)(struct rpcrdma_ia *, 380 struct rpcrdma_ep *, 381 struct rpcrdma_create_data_internal *); 382 size_t (*ro_maxpages)(struct rpcrdma_xprt *); 383 int (*ro_init)(struct rpcrdma_xprt *); 384 void (*ro_destroy)(struct rpcrdma_buffer *); 385 const char *ro_displayname; 386 }; 387 388 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops; 389 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops; 390 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops; 391 392 /* 393 * RPCRDMA transport -- encapsulates the structures above for 394 * integration with RPC. 395 * 396 * The contained structures are embedded, not pointers, 397 * for convenience. This structure need not be visible externally. 398 * 399 * It is allocated and initialized during mount, and released 400 * during unmount. 401 */ 402 struct rpcrdma_xprt { 403 struct rpc_xprt rx_xprt; 404 struct rpcrdma_ia rx_ia; 405 struct rpcrdma_ep rx_ep; 406 struct rpcrdma_buffer rx_buf; 407 struct rpcrdma_create_data_internal rx_data; 408 struct delayed_work rx_connect_worker; 409 struct rpcrdma_stats rx_stats; 410 }; 411 412 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt) 413 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data) 414 415 /* Setting this to 0 ensures interoperability with early servers. 416 * Setting this to 1 enhances certain unaligned read/write performance. 417 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */ 418 extern int xprt_rdma_pad_optimize; 419 420 /* 421 * Interface Adapter calls - xprtrdma/verbs.c 422 */ 423 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int); 424 void rpcrdma_ia_close(struct rpcrdma_ia *); 425 426 /* 427 * Endpoint calls - xprtrdma/verbs.c 428 */ 429 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *, 430 struct rpcrdma_create_data_internal *); 431 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *); 432 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *); 433 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *); 434 435 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *, 436 struct rpcrdma_req *); 437 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *, 438 struct rpcrdma_rep *); 439 440 /* 441 * Buffer calls - xprtrdma/verbs.c 442 */ 443 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *); 444 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *); 445 void rpcrdma_destroy_req(struct rpcrdma_ia *, struct rpcrdma_req *); 446 int rpcrdma_buffer_create(struct rpcrdma_xprt *); 447 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *); 448 449 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *); 450 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *); 451 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *); 452 void rpcrdma_buffer_put(struct rpcrdma_req *); 453 void rpcrdma_recv_buffer_get(struct rpcrdma_req *); 454 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *); 455 456 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *, 457 size_t, gfp_t); 458 void rpcrdma_free_regbuf(struct rpcrdma_ia *, 459 struct rpcrdma_regbuf *); 460 461 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *); 462 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int); 463 464 int frwr_alloc_recovery_wq(void); 465 void frwr_destroy_recovery_wq(void); 466 467 int rpcrdma_alloc_wq(void); 468 void rpcrdma_destroy_wq(void); 469 470 /* 471 * Wrappers for chunk registration, shared by read/write chunk code. 472 */ 473 474 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *); 475 476 static inline enum dma_data_direction 477 rpcrdma_data_dir(bool writing) 478 { 479 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE; 480 } 481 482 static inline void 483 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg, 484 enum dma_data_direction direction) 485 { 486 seg->mr_dir = direction; 487 seg->mr_dmalen = seg->mr_len; 488 489 if (seg->mr_page) 490 seg->mr_dma = ib_dma_map_page(device, 491 seg->mr_page, offset_in_page(seg->mr_offset), 492 seg->mr_dmalen, seg->mr_dir); 493 else 494 seg->mr_dma = ib_dma_map_single(device, 495 seg->mr_offset, 496 seg->mr_dmalen, seg->mr_dir); 497 498 if (ib_dma_mapping_error(device, seg->mr_dma)) 499 rpcrdma_mapping_error(seg); 500 } 501 502 static inline void 503 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg) 504 { 505 if (seg->mr_page) 506 ib_dma_unmap_page(device, 507 seg->mr_dma, seg->mr_dmalen, seg->mr_dir); 508 else 509 ib_dma_unmap_single(device, 510 seg->mr_dma, seg->mr_dmalen, seg->mr_dir); 511 } 512 513 /* 514 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c 515 */ 516 void rpcrdma_connect_worker(struct work_struct *); 517 void rpcrdma_conn_func(struct rpcrdma_ep *); 518 void rpcrdma_reply_handler(struct rpcrdma_rep *); 519 520 /* 521 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c 522 */ 523 int rpcrdma_marshal_req(struct rpc_rqst *); 524 525 /* RPC/RDMA module init - xprtrdma/transport.c 526 */ 527 extern unsigned int xprt_rdma_max_inline_read; 528 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap); 529 void xprt_rdma_free_addresses(struct rpc_xprt *xprt); 530 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq); 531 int xprt_rdma_init(void); 532 void xprt_rdma_cleanup(void); 533 534 /* Backchannel calls - xprtrdma/backchannel.c 535 */ 536 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 537 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int); 538 int xprt_rdma_bc_up(struct svc_serv *, struct net *); 539 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int); 540 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *); 541 int rpcrdma_bc_marshal_reply(struct rpc_rqst *); 542 void xprt_rdma_bc_free_rqst(struct rpc_rqst *); 543 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int); 544 #endif /* CONFIG_SUNRPC_BACKCHANNEL */ 545 546 extern struct xprt_class xprt_rdma_bc; 547 548 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */ 549